Infra-estrutura de componentes paralelos para aplicações de computação de alto desempenho

Abstract

The development of new High Performance Computing (HPC) application has demanded a set of tools for reconciling high level of a abstraction with software integration. In particular, we are interested in component-based solutions presented by the scientific community in the last years. Components have been applied to meet new requirements of high performance application such as: interoperability, reusability, maintainability and productivity. Recent approaches for component based development time in HPC context, however, have not reconciled more expressive ways for parallel programming and efficiency. Unfortunately, this issue increases the software development time and gets worse when user have poor knowledge of architectural details of parallel computers and of requirements of applications. Precious time is lost optimizing parallel code, probably with non-portable results, instead of being applied to the solutions of the problem. This dissertation presents the Hash Programming Environment (HPE), a solution based on the # (reads "Hash") Component Model and on the Hash Framework Architecture. HPE defines a set of component kinds for building, deploying and executing parallel programs targeted at clusters of multiprocessors. The Hash Framework Architecture has three loosely coupled modules: the Front-End, the Back-End and the Core. The main contribution of this work is the implementation of the Back-End and the Core. The main contribution of this work is the implementation of the Back-End, since we have an early version of the Front-End and Core, both developed in Java on top of the Eclipse Platform. The Back-End was implemented as a parallel extension of a Mono, an open source component platform based on CLI (Common Language Interface) standard. Once independently done, we bound all the modules together, using web services technology. For evaluating the proposed Back-End, we have developed a small conceptual test application, composed by # components.